from IPython.display import Audio, display, HTML
import matplotlib.pyplot as plt
import numpy as np
import os

DPI = 100
SR = 22050

Training set

def draw_example(i, wave, mel):
    fig, axes = plt.subplots(2, dpi=DPI)
    fig.suptitle('Training example %d' % (i + 1))
    axes[0].plot(wave)
    axes[1].imshow(mel.transpose(), origin='lower')
    plt.show()
    display(Audio(wave, rate=SR))

base_dir = 'mol/dump/lj/logmelspectrogram/norm/train_no_dev/'
i = 0
for fname in os.listdir(base_dir):
    path = os.path.join(base_dir, fname)
    if path.endswith('-wave.npy'):
        wave_path = path
        mel_path = path[:path.find('-wave.npy')] + '-feats.npy'
        draw_example(i, np.load(wave_path), np.load(mel_path))
        i += 1

Log probability vs time

def plot_loss_curve(title, path):
    with open(path) as f:
        losses = np.asarray([float(line.split(',')[1]) for line in f.readlines()])
    times = np.arange(len(losses))
    
    fig, ax = plt.subplots(1, dpi=DPI)
    fig.suptitle(title)
    ax.set_ylabel('Log probability')
    ax.set_xlabel('Batch')
    ax.plot(times, losses)
    plt.show()
        
plot_loss_curve('Training loss', 'mol/train_no_dev.txt')
plot_loss_curve('Test loss', 'mol/dev.txt')

Training set predictions

base_dir = 'mol/exp/lj_train_no_dev_mol_wavenet/intermediate/audio/'
for fname in sorted(os.listdir(base_dir)):
    path = os.path.join(base_dir, fname)
    if path.endswith('target.wav'):
        step = int(path[path.find('step') + 4:path.find('_target')])
        target_path = path
        pred_path = path.replace('target', 'predicted')
        
        display(HTML('<h3>Step %d</h3>' % step))
        display(HTML('<p>Training example</p>'))
        display(Audio(target_path))
        display(HTML('<p>WaveNet prediction</p>'))
        display(Audio(pred_path))

Step 1000

Training example

WaveNet prediction

Step 1200

Training example

WaveNet prediction

Step 1600

Training example

WaveNet prediction

Step 2000

Training example

WaveNet prediction

Step 2400

Training example

WaveNet prediction

Step 2800

Training example

WaveNet prediction

Step 3200

Training example

WaveNet prediction

Step 3600

Training example

WaveNet prediction

Test set predictions

base_dir = 'mol/exp/lj_train_no_dev_mol_wavenet/generated/checkpoint_latest/dev/'
for fname in sorted(os.listdir(base_dir)):
    path = os.path.join(base_dir, fname)
    if path.endswith('ref.wav'):
        target_path = path
        pred_path = path.replace('ref', 'gen')
        display(HTML('<h3>Test example %s</h3>' % path[path.find('flute'):path.find('_ref')]))
        display(HTML('<p>Ground truth</p>'))
        display(Audio(target_path))
        display(HTML('<p>WaveNet prediction</p>'))
        display(Audio(pred_path))

Test example flute_acoustic_002-075-100

Ground truth

WaveNet prediction

Test example flute_acoustic_002-077-025

Ground truth

WaveNet prediction

Test example flute_acoustic_002-077-050

Ground truth

WaveNet prediction

Test example flute_acoustic_002-077-100

Ground truth

WaveNet prediction

Test example flute_acoustic_002-078-025

Ground truth

WaveNet prediction

Test example flute_acoustic_002-078-050

Ground truth

WaveNet prediction

Test example flute_acoustic_002-078-100

Ground truth

WaveNet prediction

Test example flute_acoustic_002-079-100

Ground truth

WaveNet prediction

Test example flute_acoustic_002-079-127

Ground truth

WaveNet prediction

Test example flute_acoustic_002-080-050

Ground truth

WaveNet prediction